Re: [PATCH] Rmap speedup

[Daniel Phillips]

On Sunday 04 August 2002 02:47, Andrew Morton wrote:
> Daniel Phillips wrote:
> > 
> > On Saturday 03 August 2002 23:40, Andrew Morton wrote:
> > > - total amount of CPU time lost spinning on locks is 1%, mainly
> > >   in page_add_rmap and zap_pte_range.
> > >
> > > That's not much spintime.   The total system time with this test went
> > > from 71 seconds (2.5.26) to 88 seconds (2.5.30). (4.5 seconds per CPU)
> > > So all the time is presumably spent waiting on cachelines to come from
> > > other CPUs, or from local L2.
> > 
> > Have we tried this one:
> > 
> >  static inline unsigned rmap_lockno(pgoff_t index)
> >  {
> > -       return (index >> 4) & (ARRAY_SIZE(rmap_locks) - 1);
> > +       return (index >> 4) & (ARRAY_SIZE(rmap_locks) - 16);
> >  }
> > 
> > (which puts all the rmap spinlocks in separate cache lines)
> 
> Seems a strange way of doing it?

It is a strange way of doing it.   I felt like being engigmatic at the time, 
and no, nothing like that should ever go into production code, it would be 
better suited to an IOCCC submission.

> We'll only ever use four locks this way.

Look again: 256 - 16 = 250 = 0xf0.

-- 
Daniel